48x48 Poster Template

1
Matthew Bradley1 Kay Jantharasorn2 Keith Jones1
Advisor Dr. Mohamed Zohdy3 1 Oakland
University 2 University of Michigan- Flint
3Department of Electrical and Computer
Engineering at Oakland University
Self Organized Neural Networks Applied to Animal
Communication
UnCoRe 2008
Abstract
Feature Selection
Performance
It was found that the addition of a history to
the SOM greatly improved performance. With a
history the SOM was able to reach a stable state
within 50 iterations and the map preserved the
topological differences of the input space.
Without history, the SOM took much longer than 50
iterations(100) to reach stable state and the
overall quality was lower. Displayed below are
maps created to display the quality and topology
of the map by computing the average differences a
neuron has with its neighbors.
In this project, Self Organizing Feature Maps
are trained to categorize animal communication
sounds into danger, hunger, and mating calls for
Humpback Whales, Bottlenose Dolphins and Coyotes.
Features are extracted in time domain, frequency
domain, and joint time-frequency domain from
audio files of animal communication. Unknown
calls are then fed into the map to identify the
sound. Several contributions have been made to
the maps to improve the quality, speed, and
accuracy. Correlation has been implemented as
the activation function to improve the accuracy
of identifying sounds. A history function to
update previous winning nodes has been
implemented to decrease the time necessary to
complete the training phase. Sammon mapping has
also been proposed to provide better
initialization of vectors. In addition, several
types of distance metrics are tested to find best
matching units including a blended
Manhattan-Euclidean distance which has been found
useful.
W e extracted data from frequency domain, time
domain and joint time-frequency domain of each
sound from each animal. We filtered frequency
domain to minimize noise. We selected several
features from time and frequency domains as
vectors and joint time frequency domain is
selected as a matrix.
Frequency Domain Frequency
Domain Time Domain
Time/Frequency
Filtered
Background
A Self-Organizing Feature Map(SOFM) is a
type of unsupervised neural network designed to
map high dimensional spaces onto low dimensional
spaces that can be easily understood and
visualized. The map consists of input nodes to
which feature vectors of high dimension are
presented. SOFMs are trained on exemplary
patterns using two stages competitive and weight
update stage.
With History Without History
When testing the accuracy of the maps it was
found that the SOFM that used features from the
frequency domain performed better overall at
approximately 74.6, followed by the
time-frequency domain at 63.5, and lastly the
time-domain at 55.5 accuracy.
Distance Selection
Wij-- Neuron Weights a(t)- Learning Rate
ß(t)- Neighborhood Function X -- Input
Vector
Objective
Conclusion
There are several different ways to compute
distance(Shown to the right) 1-Norm(Manhattan) 2-
Norm(Euclidean) p-norm Infinity
In conclusion, self-organizing feature maps were
successful in categorizing and identifying
danger, hunger, and mating calls of Coyotes,
Humpback Whales, and Bottlenose dolphins. The
history function greatly improves the performance
during the training phase. Correlation increases
the success rate of identifying unknown sounds by
15-20 Blended distance not only has a
significant effect in improvement of
computational time in our self-organizing feature
maps but it can also be used in a number of
applications. We are currently writing three
papers to be submitted for publication and
conferences. The main paper will generalize our
findings and be submitted to IEEE, while the
other papers will specialize in either SOFM
contributions or animal communication and be
submitted to conferences.

• The main objective of this project is
  to extract features from several animal sounds
  that were communicated under situations of
  mating, danger, and foraging using MATLAB. An
  application was developed in Java to train
  several SOFMs using those data. This allows us
  to discern the animal sounds and the situation
  under which theyre used. We have made the
  following contributions in this project
• On Self-Organizing Feature Map
• Input Presentation Order
• Winning neurons history
• Taking account of context in which the sounds are
  made
• Proposing Correlation to do finer level
  clustering
• Sammons projection introduced to provide better
  initialization to the SOFM
• On Feature Selection Ordering
• Proposing Matrix Features in the case of joint
  time-frequency domain
• Using normalization of input data for better
  efficiency
• Pre-processing the input to remove noise
• On Animal Sounds
• Humpback Whale
• Bottle nose dolphin
• Coyote

Metrics between Euclidean and Manhattan are
useful to obtain best matching units when the
data has a rotational bias. A blended
Euclidean-Manhattan distance metric (lambda) is
proposed to approximate the traditional Lp metric
(for p 1 to 2) while costing considerably less
computational time
Acknowledgement s
Wed like to thank National Science Foundation
and Oakland University for giving us this great
opportunity to explore and participate in
research and learn more about research careers.
Wed also like to thank our advisor, Dr. Mohamed
Zohdy for his guidance throughout this research,
Doug Hunter for his scientific journals, articles
related to Humpback whales, and Humpback whales
data.
References
Kohonen, T., Self-Organizating Maps,  New York
Springer- Verlag, 1997 Payne, Roger S. McVay,
Scott. Songs of Humpback Whales. Science 173
(1971) 585-597 Germano, Tom. 23 March, 1999. 2
June, 2008. lthttp//davis.wpi.edu/matt/courses/s
oms/gt Hunter, Doug. Professor Emeritus. Biology.
Oakland University Nsour, Ahmad. Zohdy,
Mohamed. Self Organized Learning Applied to
Global Positioning System (GPS) Data. Oakland
University FindSounds. 2008. Comparisonics. 2
June, 2008. lthttp//www.findsounds.com/gt